| name | researcher |
| description | Meta-skill that keeps all agents, skills, and knowledge files current and honest — Daily Quick Check + Weekly Deep Research modes. Categorization schema (Tier 1 Consensus / 2 Strong-but-Contextual / 3 Divergent / 4 Emerging / 5 Deprecated). Includes 90-day staleness checks for decision trees AND `.ravenclaude/environment-context.md`. |
Researcher Skill (Meta-Skill)
Purpose
The Researcher is a meta-skill responsible for keeping all agents, skills, and knowledge files in the RavenClaude marketplace current and honest.
It runs in two modes:
- Daily Quick Check — Recommended the first time the repo is opened each day (or manually invoked by Team Lead / Grok Captain).
- Weekly Deep Research — Comprehensive review of every agent + its associated skills and knowledge (recommended weekly, e.g. Sunday/Monday).
How to Trigger
Daily Quick Check:
- Team Lead (or Grok when acting as orchestrator) should run a lightweight version on first meaningful work session of the day.
- Focus: Quick scan for any obviously outdated advice in active agents/knowledge areas.
Weekly Deep Research:
- Full sweep across all agents and knowledge files.
- Best done as a dedicated session.
- Can be triggered manually or via scheduled reminder.
Core Principles
- Grounding First: Before updating anything, apply the Capability Grounding Protocol.
- Consensus + Divergence: Capture both the widely accepted expert view and credible dissenting views.
- Actionable Updates: Research must result in specific, justified proposals to update agents, skills, or knowledge files.
- Transparency: Every research output must clearly label information as Consensus, Divergent, Emerging, or Contextual.
Research Scope
For each agent:
- Review its current definition and responsibilities.
- Identify all skills and knowledge files it relies on.
- Research recent changes, best practices, gotchas, and expert opinions (official + community + divergent).
- Categorize new information using the schema.
- Propose concrete updates with justification.
Categorization Schema (Mandatory)
All researched information must be categorized:
Tier 1: Consensus / Widely Accepted (Default)
Backed by official Microsoft documentation + strong agreement among recognized experts and MVPs.
Tier 2: Strong but Contextual
Generally recommended but has known limitations or scenario-specific caveats.
Tier 3: Divergent / Contrarian Views (Critical Fallback)
Credible experts who successfully do things differently. Include their reasoning and when their approach may be better.
Tier 4: Emerging / Experimental
New patterns or preview features with early positive signals.
Tier 5: Deprecated or Risky
Previously common approaches now discouraged.
Output
Produce a structured Research Report using templates/research-report-template.md.
The Researcher must apply the Grounding Protocol to its own conclusions before presenting updates.
Sources
Official: Microsoft Learn, release notes, Microsoft 365 Roadmap, official samples.
Community: High-signal MVPs, forums, quality blogs/YouTube.
Divergent: Practitioners who publicly challenge common advice with demonstrated results.
Actively seek credible dissenting views rather than only confirming existing knowledge.
Learn-tab improvement pass (added 2026-06-12)
The Weekly Deep Research sweep for Tier-A web-design and frontend-engineering plugins also feeds a Learn-tab improvement pass — checking whether new platform facts, documentation patterns, or web-design / interactivity best practices warrant a new concept card (or an update to an existing one) in plugins/ravenclaude-core/knowledge/concepts/.
Domains to sweep:
- Documentation best practices — Diátaxis (diataxis.fr), Google developer documentation style guide, Microsoft Writing Style Guide, Write the Docs newsletter / community updates. Look for: new structural frameworks (tutorials/how-tos/reference/explanation split evolution), updated style conventions, tooling changes (MkDocs/Docusaurus/Sphinx releases).
- Website design best practices — web.dev/blog, MDN Web Docs "What's New", Chrome Developers / Safari WebKit release notes, W3C/WHATWG spec changelogs for active proposals. Look for: new CSS/layout primitives (container queries, scroll-driven animations, view transitions, anchor positioning), new browser APIs, updated Core Web Vitals thresholds, accessibility standard updates (WCAG 3.x progress).
- Interactivity best practices — Google Web Vitals guidance (INP/LCP/CLS thresholds and measurement), ARIA Authoring Practices Guide (APG) updates, progressive enhancement patterns, motion design (
prefers-reduced-motion), Web Animations API, keyboard/focus management patterns.
Primary sources:
- web.dev/blog and developer.chrome.com (weekly platform signals)
- MDN What's New (per browser release)
- WebKit/Safari release notes and WWDC talks (annual + point releases)
- W3C/WHATWG spec changelogs for active proposals
- Nielsen Norman Group research publications (UX/interactivity patterns)
- Diátaxis.fr and Write the Docs newsletter (documentation frameworks)
Output — for each notable finding, propose one of:
- A new concept card in
knowledge/concepts/ (with full + mini Mermaid diagram) if the concept is broadly applicable to documentation or web design work in any consumer project.
- An update to the relevant knowledge file in
plugins/web-design/knowledge/ or plugins/technical-writing-docs/knowledge/.
- An honest null result ("0 net-new this week") — logged, not padded.
Authoring guard: concept cards require a mermaid-cli render pass (render-concepts.py). Propose concept content in a staging doc or PR description; do not write a bare .md without an accompanying SVG unless the rendering pipeline is available in the session.
Environment-context staleness check (added 2026-05-22)
The Weekly Deep Research sweep MUST also check the consumer's .ravenclaude/environment-context.md (if present) for staleness:
- Read the file's
Last reviewed: field (or git last-modified if no field exists)
- If older than 90 days, surface in the Research Report as action: re-run
environment-discovery or manual refresh
- The file silently going stale produces the failure mode this whole mechanism was built to prevent — agent forgets it's authorized and starts asking the user for auth again. The 90-day check is the anti-stale backstop.
This check runs alongside the decision-tree staleness check below; both feed the same "priors aging out" concern.
Decision-tree staleness check (added 2026-05-21)
The Weekly Deep Research sweep MUST include a staleness check for decision trees per the convention in ../../../docs/best-practices/decision-trees-in-knowledge-files.md.
The check:
- Glob
plugins/*/knowledge/*.md and plugins/*/skills/*.md for files containing ## Decision Tree: section headers
- Parse the
**Last verified:** YYYY-MM-DD field within each
- Flag any tree where
today - last_verified > 90 days
- For each flagged tree, run the categorization schema (Consensus / Contextual / Divergent / Emerging / Deprecated) against each leaf — has any leaf become inaccurate since the last verification?
- Surface flagged trees in the Research Report with one of:
still-current (refresh date only), needs-update (specific leaves to revise), deprecate (remove tree entirely).
Why this check matters: decision trees go stale faster than prose when underlying platforms change. A prose paragraph degrades gracefully ("the API used to return X, now Y" still reads fine). A decision tree with a 404 → reimport leaf is wrong the moment the platform returns 409 instead. The Last verified: field + this check is the anti-staleness backstop the format requires.